Colon polyp segmentation method based on global contextual information and U-shaped network

In response to the challenges present in colon polyp image segmentation such as large-scale variations in lesion areas, blurred edges, and low contrast between polyps and normal tissues, traditional convolutional neural networks lead to loss of detailed information and inadequate interaction between features at different levels, consequently decreasing segmentation accuracy. This study proposes an enhanced network named GCI-UNet (Improved network based on global context in-formation and U shaped network). The network utilizes Swin Transformer encoders to progressively extract global contextual information from input images, attenuating background interference, and conducting multiscale analysis of significant features in lesion areas. Additionally, by leveraging Skip Connections in U-shaped networks, it integrates detailed information extracted from shallow networks and abstract information from deep networks, thereby enhancing seg-mentation outcomes by consolidating edge details. Furthermore, the network’s spatial awareness is enhanced through a global and local feature fusion module, emphasizing critical positional in-formation of segmentation targets and achieving precise localization, thereby further enhancing segmentation accuracy. Experiments on the publicly available CVC-ClinicDB dataset demonstrate that GCI-UNet achieves a Dice coefficient of 0.945 and an MIoU of 0.898, indicating significantly superior segmentation performance compared to existing methods. Comparative experimental results reveal that GCI-UNet exhibits higher segmentation accuracy when handling complex morphology, low contrast, and blurred edges in colon polyp images.